Or well… almost. It’s in the hands of various review sites, and due to be available through retail channels in about 2 weeks. However, thanks to these review sites, we already have a reasonably good idea of what to expect.
The short version is that the GTX480 is about 15% faster than the Radeon 5870 on average. Given the higher price, heat, noise and power consumption, this doesn’t quite make the GTX480 a compelling option for the average user. However, I am not an average user…
What I found especially interesting were the cutting-edge features of Fermi, such as the new parallelized tessellation approach, and the improved GPGPU capabilities. Various reviews used the Unigine Heaven benchmark or DirectX 11 samples to demonstrate tessellation performance, and this is one area where Fermi really shines. Where the Radeon gets choked quite quickly, Fermi is able to scale through high tessellation factors quite well. Tessellation is only interesting when you can apply it on a large scale with virtually no performance hit. Otherwise you might be better off with just using more detailed geometry directly, or using an approach such as parallax occlusion mapping, giving similar detail with better performance. For the same reason, the geometry shader in DX10 never really took off. It just was too slow to be useful for realtime tessellation. It’s hard to tell at this point whether Fermi’s increased tessellation performance is enough to really make tessellation a success, but at least it looks considerably better than AMD’s effort.
On the GPGPU side, Fermi seems to be an absolute monster. In various tests it seems to be more than twice as fast as the GTX285, which was no slouch in GPGPU itself. In Folding@Home it even manages to be 3.5x as fast as the GTX285. This compute power also gives PhysX a healthy boost. Another interesting thing was the nVidia Design Garage raytracing demo. This got a tremendous boost as well, which is a testament Fermi’s massively improved caching/memory access, and demonstrates that Fermi takes GPGPU one step closer towards conventional CPUs in that sense.
In terms of image quality, there doesn’t seem to be much to talk about. The antialiasing and texture filtering appear to be pretty much identical to the previous generation of nVidia GPUs. They have added a new 32xAA mode however, although I’m not sure how useful that will be in practice. This isn’t necessarily a bad thing, as the previous generation had excellent AA and AF quality already… It however does not take that final step to ‘perfection’ that AMD had made with the Radeon 5000 series, using ‘perfect’ angle-independent AF (which I personally don’t use, as I run with Catalyst AI enabled, so AF quality is sacrificed for speed… but you can’t really tell anyway).
My conclusion so far? The Fermi architecture has a lot of potential for the future, with its good tessellation and GPGPU performance… However, it seems it hasn’t quite reached its sweet-spot yet. I personally find them a bit too power-hungry. Perhaps TSMC can get the power consumption down over the next few months as the production process matures, and perhaps with small bumps in clockspeed and driver improvements, the power/performance ratio can become more acceptable. Or perhaps we need to wait for a die-shrink to get there.
I would have considered the GTX470 for myself… but I’m not too comfortable with its power consumption. So I think I’ll give it a few months to see how Fermi matures. If TSMC can get power consumption down, if nVidia can get the performance up via the drivers, and if the retail price is attractive enough, then I will buy one. But the first impression is not good enough for me to pre-order a card right away.
It could also be that one of the ‘stripped down’ variations of the Fermi architecture will turn out to be a good deal. It could be that a smaller die size will have such a favourable effect on yields that it turns performance and power consumption around completely. That the full-blown Fermi is a bit too ‘obese’, but a leaner and meaner variation of the architecture does hit that sweet-spot.